Author: Daniele Di Mitri

Daniele Di Mitri is a professor of Multimodal Learning Technologies at the German University of Digital Science. At the German UDS, he leads the research group "Augmented Feedback" and coordinates the master's in Advanced Digital Realities. He is an associated researcher at the DIPF - Leibniz Institute for Research and Information in Education and a lecturer at the Goethe University of Frankfurt, Germany. Daniele Di Mitri received his PhD in Learning Analytics and Wearable Sensor Support from the Open University of the Netherlands. His current research focuses on developing AI-driven, multimodal learning technologies to enhance digital education. It aims to create innovative, responsible solutions that improve learning experiences through advanced feedback systems and ethical integration of technology. He is a "Johanna Quandt Young Academy" fellow and was elected "AI Newcomer 2021" at the KI Camp by the German Informatics Society. He is a member of the CrossMMLA, a special interest group of the Society for Learning Analytics Research, and the chair of the special interest group on AI for Education of the European Association for Technology-Enhanced Learning.

Conference article

New pub: Structured Over Chat-Based Feedback

Posted on 10/13/2026by Daniele Di Mitri

The paper, Structured over Chat-Based Feedback: Comparing User Preferences on Different Presentation Feedback Reports, by Noemi Giulia Bontempo, Nina Mouhammad and Daniele Di Mitri, explores how learners prefer to review feedback after practising a presentation, focusing on a question that receives less attention than immediate coaching: how should feedback be presented once the rehearsal is over?

Digital presentation-training systems can provide feedback on delivery behaviours such as posture, gestures and orientation towards the camera or audience. But during a presentation, learners are already managing their content and delivery. A report after the rehearsal offers another opportunity: to revisit performance and reflect on what to improve. We compared three ways to support that process in Presentable: a timeline that links feedback to moments in the presentation, a chat-based conversation with a virtual coach, and a storytelling-style report organised into themes such as gestures, posture and camera connection.

We evaluated the three designs in an online study with 84 participants. Everyone reviewed all three versions, which used the same fictional rehearsal scenario and feedback content. This allowed us to compare the presentation of feedback rather than differences in the feedback itself. Participants rated each design and compared its perceived clarity, usefulness, support for reflection and potential to help them understand their performance.

The two structured designs were rated more positively overall than the chat-based report. The timeline was selected as most helpful overall by 48% of participants, while 40% chose the storytelling design and 12% chose the chat design. The timeline was particularly valued for showing when behaviours occurred. The storytelling design was selected most often as easiest for understanding presentation performance, by 49% of participants.

Participants also highlighted the design elements they found useful for reflection: clear visual grouping of categories (73%), highlighting key moments (61%), and making the connection between feedback and time clear (55%). Some appreciated the chat’s personal and interactive feel, but others found it harder to get an overview, compare feedback points or return to specific items.

For us, the findings suggest that conversational presentation alone does not necessarily make feedback easier to use. When learners review a rehearsal, clear organisation and context matter: they need to see what happened, when it happened and how different observations fit together. The results also point towards a promising next step: combining the timeline’s moment-by-moment detail with the storytelling design’s readable thematic overview.

We interpret these results as design insights, not evidence that one report format improves presentation skills. Participants assessed mock-ups based on a fictional scenario; they did not use the reports after delivering their own presentations. Our next step is to integrate the Rehearsal Report into Presentable and connect feedback items to the corresponding moments in a learner’s recording, enabling evaluation in real practice sessions.

Citation:

Bontempo, N. G., Mouhammad, N., & Di Mitri, D. (2026). Structured over Chat-Based Feedback: Comparing User Preferences on Different Presentation Feedback Reports. European Conference on E-Learning, 25(1), 99–108. https://doi.org/10.34190/ecel.25.1.5389

Conference article

New pub: Educator as Co-pilot: Advancing Presentation Training with Teacher-guided Systems

Posted on 10/09/2026by Daniele Di Mitri

The newly published paper Educator as Co-pilot: Advancing Presentation Training with Teacher-guided Systems. grew out of Talha Burak Onat’s master’s thesis at Goethe University Frankfurt, conducted in collaboration with DIPF | Leibniz Institute for Research and Information in Education.

Presentation-training platforms can give students opportunities to rehearse and receive automated feedback on their verbal and non-verbal communication. But automated analysis cannot always account for context or the learning goals of a particular presentation. In this work, we wanted to explore how teachers could add that human perspective: interpreting feedback, offering guidance, and helping students decide what to work on next. This concept has since shaped the current approach used by Presentable (www.presentable.info), where automated analysis is complemented by teacher guidance.

To investigate this, we designed and evaluated a teacher-feedback feature for Presentable, a web-based presentation-training platform. It enables educators to write structured comments using templates, link feedback to specific moments in a video, annotate transcript passages, and validate or reject automated feedback before students see it. Students and teachers can also exchange replies, making feedback an ongoing conversation rather than a one-way report.

We explored the tool through interviews with four educators: two secondary-school teachers and two university-level tutors. They generally found the interface clear and valued the templates, timestamp-linked comments and transcript annotations. These functions helped connect feedback to specific parts of a presentation, while written comments and asynchronous discussion offered ways for students to revisit and clarify advice.

The interviews also showed us where the design needs further work. Educators asked for draft-saving and autosave, clearer access to transcript comments, and the ability to edit submitted feedback. Their views on automated feedback were mixed: some messages drew attention to presentation behaviours, while others felt trivial or unrelated to assessment criteria. This reinforced an important point for us: automated feedback needs to be relevant and useful in context, and teachers need the means to review it.

We see the paper as an early step towards presentation-training systems in which automation supports, rather than replaces, educators’ expertise. Our evaluation involved a small group, so it offers initial design insights, not evidence of improved student performance. Further studies with larger and more varied groups will be needed to examine how teacher-guided feedback affects learning over time.

We are especially glad that this publication brings work from Burak’s master’s thesis into a broader conversation about human-centred educational technology. It shows how a thesis project can move from identifying a practical gap, through designing and evaluating a solution, to informing the next stage of a research platform.

Citation:

Di Mitri, D., Onat, T.B. (2026). Educator as Co-pilot: Advancing Presentation Training with Teacher-Guided Systems. In: Adamoli, M., et al. Higher Education Learning Methodologies and Technologies Online. HELMeTO 2025. Communications in Computer and Information Science, vol 2958. Springer, Cham. https://doi.org/10.1007/978-3-032-31850-3_28

Conferences

Reflections about LAK26 and where the field is heading

Posted on 05/04/2026by Daniele Di Mitri

I was lucky this week to attend the Learning Analytics '26 conference in Bergen, Norway. This year's conference focused on the synergies between LA and Generative AI. This shift to GenAI has intensified in the last few years. The collection of data from more traditional sources, such as LMS logs or visualisations in the LA dashboard, has been replaced by efforts to capture how and to what extent students learn with GenAI.

This shift is also reflected in the workshop topics. For instance, in the CROSSMMLA workshop, we explored GenAI as a "sensor for semantics" that can be integrated with a variety of modalities to analyse the learning process and add a layer of deeper understanding to typically structured and messy multimodal data.

Since my first LAK in 2016, I have been eagerly following the development of the field, while generally being quite positive yet critical of the research community's openness to new, theory-informed, technically rich approaches.

This year, however, the progressive shift towards GenAI at LAK left me not with enthusiasm but with a sense of unsettlement about the field.

First of all, there is the realisation that scientific discourse has pivoted almost exclusively toward how to make LLMs work for specific educational purposes, regardless of whether they are suitable or convenient to use over more parsimonious approaches. This includes how to train, fine-tune, and, more generally, "tame" LLMs, as well as how to deal with their side effects, such as fabricated results and incorrect information.

But very few of these works have addressed why these systems should be used in the first place, nor have they explored the broader consequences of using LLMs, e.g., resource exploitation, data labour by underpaid workers, and copyright infringement.

The dominant scientific imperative is to use LLMs as a research method, regardless of the results they produce, whether their use offers an actual advantage for students, learners, or a more powerful scientific approach.

It seems to me that science is also a victim of the hype rhetoric that either uses GenAI or is left behind. It is sad but true to admit that LA research is slowly being swept away by GenAI.

The critique of GenAI and the economy of hyperscale is probably an ethical dilemma I see, while many fellow scientists don't see it as an ethical problem at all.

Adapting to GenAI is imperative in the current era, where LLM use is pervasive, and adoption is unprecedented. While I see that this technology is here to stay, I am not blindly buying it, and I believe that researchers cannot absolve themselves of the responsibility to examine the social ramifications of a technology, just because it is widely used.

There is no straightforward positioning here. If we do not want to be swept away even more by GenAI and the corporations behind it, we have to strengthen our critical thinking skills and question how and why we do things, as well as the net advantages.

Conference article

New pub: Are rubrics all you need? Towards rubric-based automatic short answer scoring

Posted on 04/27/2026by Daniele Di Mitri

The latest paper led by Sebastian Gombert has been published in the Proceedings of the LAK26: 16th International Learning Analytics and Knowledge Conference (LAK26).

"Are rubrics all you need? Towards rubric-based automatic short answer scoring via guided rubric-answer alignment"

In educational assessment, rubrics are central because they define clear criteria for evaluating learner responses and specify what counts as relevant evidence. Yet, most automatic short answer scoring approaches make little to no explicit use of rubrics, or treat them only as additional side information. This paper turns that around and asks what happens if rubrics themselves become the primary scoring reference for automated systems.

The authors introduce the task of rubric-based automatic short-answer scoring, in which the model uses the scoring rubric as an explicit anchor rather than relying solely on large sets of labelled student responses. To implement this idea, they propose a guided rubric–answer alignment, in which each student's answer is aligned directly with rubric criteria and level descriptors rather than with other answers.

Building on this concept, the paper presents two new transformer-based architectures, GRAASP and ToLeGRAA, which use attention mechanisms to focus on the most relevant rubric information when predicting scores. These architectures aim to make scoring more transparent and more faithful to the assessment design, and they promise greater robustness when tasks change because the scoring logic is driven by the rubric rather than solely by historical training data.

This work aligns with a broader agenda in our group: designing AI systems that are tightly coupled with pedagogical artefacts such as rubrics, feedback guidelines, and learning objectives, instead of treating AI as a detached black box. By placing rubrics at the centre of the modelling process, this research opens a path towards more interpretable, educator-aligned automatic assessment tools that can better support teaching and learning.

Check it here (Open Access PDF via ACM):

Gombert, S., Sun, Z., Zehner, F., Lossjew, J., Wyrwich, T., Czinczel, B. K., Bednorz, D., Kubsch, M., Di Mitri, D., Neumann, K., & Drachsler, H. (2026). Are rubrics all you need? Towards rubric-based automatic short answer scoring via guided rubric-answer alignment. Proceedings of the LAK26: 16th International Learning Analytics and Knowledge Conference, 272–282. https://dl.acm.org/doi/10.1145/3785022.3785064

Journal article

New pub: Through the Telescope: A Systematic Review of Intelligent Tutoring Systems

Posted on 12/08/2025by Daniele Di Mitri

The systematic literature review led by Gianluca Romano has been published in the International Journal of Artificial Intelligence in Education by Springer Nature

"Through the Telescope: A Systematic Review of Intelligent Tutoring Systems and Their Applications in Psychomotor Skill Learning"

This review fits in with our broader effort as a group on how AI can be supportive for psychomotor skills, i.e. those skills which require mind-body coordination, and that have a high degree of physicality.

The article systematically reviews "Intelligent Tutoring Systems (ITS)" and finds that current ITS primarily support fine, simple, and technical skills, such as those in medical and sports training.

We highlight gaps in addressing complex, gross, and open skills. For the future of the field, we call for ITS to incorporate broader physical skill dimensions, personalised feedback, and training theories to achieve more effective, holistic skill development. In the future, we expect ITS to move beyond repetition and expert comparison toward adaptive, theory-driven learning support.

Check it here Open Access 🔓

Romano, G., Schneider, J., Di Mitri, D. et al. Through the Telescope: A Systematic Review of Intelligent Tutoring Systems and Their Applications in Psychomotor Skill Learning. Int J Artif Intell Educ (2025). https://link.springer.com/article/10.1007/s40593-025-00526-1